4 ms·

You might actually expand "Apple likes open standards when it is weak" to "companies likes open standards when they are weak." Generally speaking you get stand

You might actually expand "Apple likes open standards when it is weak" to "companies likes open standards when they are weak." Generally speaking you get standards consortiums when there is a clear winner that is mopping up the space. Here's an example that's happening right now: Nvidia-NvLink-Infiniband. Nvidia owns the highspeed interconnect inside the chassis (HGX), the NICs (Mellanox), the inter-host interconnect (Infiniband), the high performance inter-host interconnect (switched NvLink), the and the ethernet network (Mellanox has the same 52.1Tbps switch performance that everyone else has now). GPU training is RDMA heavy and this is a place where both NvLink and Infiniband shine, ethernet much less so. Retransmissions are very bad, in global-performance-terms, for ROCEv2 transfers. Right now Nvidia is just crushing it and there's zero chance anyone is going to catch up by introducing new Infiniband ASICs. So what happens? You have a consortium spun up by all of the companies in the Ethernet space - Ultra Ethernet Consortium - to try and use "standards" to push back on customers who don't want to make big investments in "non-standard." UEC is pretty vague but seems to be promising Broadcom-style cellized fabrics, the whole point of which is to have an ethernet-like standard that avoids ECMP-induced flow collisions and retransmissions - that is, get Ethernet into the same territory as Infiniband. If you look back in time in the tech industry, you see this over and over and over and over. Standards are great, they make certain kinds of multi-sided markets and markets that need broad participation to be viable possible - but they are also routinely about the losers joining together to compete.

paulmd 3y ago
Same for adaptive sync. Gsync was first to market by a country mile, the adaptive sync standard wasn’t even approved until like 6 months after the first gsync stuff showed up and if nvidia hadn’t gone first it would have taken much longer if it happened at all. Freesync wasn’t really market-ready until at least 2015 and the early products were mostly junk, it’s not until the gsync compatible program that any vendors really cared about LFC or flickering issues.
SergeAx 3y ago
> losers joining together to compete You are saying it like it is something bad. Competition is good for consumers, and I cheer any means to spin up, sustain and heat competition.
galaxyLogic 3y ago
> they are also routinely about the losers joining together to compete. That is the greatness of it. It reminds me of democracy: The less powerful join together to give everyone an equal vote, rather than having one vote per dollar.
- robertlagrant 3y ago
  In this case though dollars work just fine - the incumbent who probably invented the market now gets smaller companies banding together against it after a while.
- realjhol 3y ago
  [dead]
michh 3y ago
> Right now Nvidia is just crushing it and there's zero chance anyone is going to catch up by introducing new Infiniband ASICs. I mean, Cornelis Networks is trying with the resurrected OmniPath. I hope they pull it off, but I'm not holding my breath. See: https://www.cornelisnetworks.com/products/ https://www.cornelisnetworks.com/products/
- foobiekr 3y ago
  It's not going to work because they are too late. HPC cluster builds are complex enough due to the presence of multiple networks (2x moderate-scale infiniband chassis and 2x ethernet chassis as a minimum) without introducing unknown vendors. At that point, if you're doing IB, why not just go Mellanox since you will almost certainly buy 200GE Connectix NICs and not the 100G NICs from these guys. UEC will - like most standards - _eventually_ work. In the mean time, Nvidia pods are the obvious choice for anyone who really cares about performance, and other vendors (Cisco, Arista) if they don't.
  - michh 3y ago
    Idk, I personally did it for a HPC cluster two years ago. 2x100GbE + OmniPath was a sensible way to reduce cost, especially as the cluster was very light on GPU power and mostly focussed on CPU-bound jobs. Last I heard everyone there is still very happy with what we built.
    - foobiekr 3y ago
      Was that before Intel spun it out? I can see people being willing to do that build if Intel was seemingly on board. Today things are pretty different. Modern HPC mostly means GPU compute and tons of data shuffling, but point taken. CPU bound jobs aren't going to stress the i/o, so you probably could have done 100GE for less. I'm curious what you did for storage but I'm guessing with CPU-bound that is again much less of an issue.
krsdcbl 3y ago
I'd even bend that quote to say "Given a problem that maximizes their profit if solved, it's an acceptable compromise for companies to resort to solutions they don't control, if they don't have the ressources to create competing solutions they'd do control."
nixpulvis 3y ago
While we’re over generalizing… What could be said about grand unification vs the march of progress? Can the laws of physics change? Must we reinvent arithmetic? One of my greatest joys learning to program is reading old (well written) code, and piecing apart the timeless from the legacy.
crest 3y ago
Or wait for the incumbent to get big fat and lazy before beating them using a 10x faster, dumb, "unsuitable" interconnect loaded to only to half its theoretical bandwidth.
- sassy_quat 3y ago
  [dead]

paulmd 3y ago

Same for adaptive sync. Gsync was first to market by a country mile, the adaptive sync standard wasn’t even approved until like 6 months after the first gsync stuff showed up and if nvidia hadn’t gone first it would have taken much longer if it happened at all. Freesync wasn’t really market-ready until at least 2015 and the early products were mostly junk, it’s not until the gsync compatible program that any vendors really cared about LFC or flickering issues.

SergeAx 3y ago

> losers joining together to compete You are saying it like it is something bad. Competition is good for consumers, and I cheer any means to spin up, sustain and heat competition.

galaxyLogic 3y ago

> they are also routinely about the losers joining together to compete. That is the greatness of it. It reminds me of democracy: The less powerful join together to give everyone an equal vote, rather than having one vote per dollar.

robertlagrant 3y ago
In this case though dollars work just fine - the incumbent who probably invented the market now gets smaller companies banding together against it after a while.
realjhol 3y ago
[dead]

michh 3y ago

> Right now Nvidia is just crushing it and there's zero chance anyone is going to catch up by introducing new Infiniband ASICs. I mean, Cornelis Networks is trying with the resurrected OmniPath. I hope they pull it off, but I'm not holding my breath. See: https://www.cornelisnetworks.com/products/ https://www.cornelisnetworks.com/products/

foobiekr 3y ago
It's not going to work because they are too late. HPC cluster builds are complex enough due to the presence of multiple networks (2x moderate-scale infiniband chassis and 2x ethernet chassis as a minimum) without introducing unknown vendors. At that point, if you're doing IB, why not just go Mellanox since you will almost certainly buy 200GE Connectix NICs and not the 100G NICs from these guys. UEC will - like most standards - _eventually_ work. In the mean time, Nvidia pods are the obvious choice for anyone who really cares about performance, and other vendors (Cisco, Arista) if they don't.
- michh 3y ago
  Idk, I personally did it for a HPC cluster two years ago. 2x100GbE + OmniPath was a sensible way to reduce cost, especially as the cluster was very light on GPU power and mostly focussed on CPU-bound jobs. Last I heard everyone there is still very happy with what we built.
  - foobiekr 3y ago
    Was that before Intel spun it out? I can see people being willing to do that build if Intel was seemingly on board. Today things are pretty different. Modern HPC mostly means GPU compute and tons of data shuffling, but point taken. CPU bound jobs aren't going to stress the i/o, so you probably could have done 100GE for less. I'm curious what you did for storage but I'm guessing with CPU-bound that is again much less of an issue.

krsdcbl 3y ago

I'd even bend that quote to say "Given a problem that maximizes their profit if solved, it's an acceptable compromise for companies to resort to solutions they don't control, if they don't have the ressources to create competing solutions they'd do control."

nixpulvis 3y ago

While we’re over generalizing… What could be said about grand unification vs the march of progress? Can the laws of physics change? Must we reinvent arithmetic? One of my greatest joys learning to program is reading old (well written) code, and piecing apart the timeless from the legacy.

crest 3y ago

Or wait for the incumbent to get big fat and lazy before beating them using a 10x faster, dumb, "unsuitable" interconnect loaded to only to half its theoretical bandwidth.

sassy_quat 3y ago
[dead]